A Multi-Level Constraint-Based Controller for the Dynamo98 Robot Soccer Team

Constraint Nets provide a semantic model for modeling hybrid dynamic systems. Controllers are embedded constraint solvers that solve constraints in real-time. A controller for our new softbot soccer team, UBC Dynamo98, has been modeled in Constraint Nets, and implemented in Java, using the Java Beans architecture. An evolutionary algorithm is designed and implemented to adjust the weights of constraints in the controller. The paper demonstrates that the formal Constraint Net approach is a practical tool for designing and implementing controllers for robots in multi-agent real-time environments. 1 Background and Introduction Soccer as a task domain is su ciently rich to support research integrating many branches of robotics and AI [3, 6]. To satisfy the need for a common environment, the Soccer Server was developed by Noda Itsuki [1] to make it possible to compare various algorithms for multi-agent systems. Because the physical abilities of the players are all identical in the server, individual and team strategies are the focus of comparison. The Soccer Server is used by many researchers and has been chosen as the o cial simulator for the RoboCup Simulation League [2]. Constraint Nets (CN), a semantic model for hybrid dynamic systems, can be used to develop a robotic system, analyze its behavior and understand its underlying physics [8{10]. CN is an abstraction and generalization of data ow networks. Any (causal) system with discrete/continuous time, discrete/continuous (state) variables, and asynchronous/synchronous event structures can be modeled. Furthermore, a system can be modeled hierarchically using aggregation operators; the dynamics of the environment as well as the dynamics of the plant and the controller can be modeled individually and then integrated [7]. A controller for our new softbot soccer team, UBC Dynamo98, has been developed using CN. The rest of the paper describes CN and how we use it to model and build the controller for our soccer-playing softbot UBC Dynamo98. Section 2 introduces the CN model of the controller for our soccer-playing softbot. Section 3 discusses constraint-based control and shows how the controller satis es the constraints M. Asada and H. Kitano (Eds.): RoboCup-98, LNAI 1604, pp. 402-409, 1999. c Springer-Verlag Heidelberg Berlin 1999 in the soccer domain. Section 4 shows our team's performance in RoboCup98. Section 5 concludes the paper. 2 The CN Architecture of the Controller for a Soccer-playing Softbot 2.1 Modeling in Constraint Nets A constraint net consists of a nite set of locations, a nite set of transductions and a nite set of connections. Formally, a constraint net is a triple CN = hLc; Td;Cni, where Lc is a nite set of locations, Td is a nite set of labels of transductions, each with an output port and a set of input ports, Cn is a set of connections between locations and ports. A location can be regarded as a wire, a channel, a variable, or a memory cell. Each transduction is a causal mapping from inputs to outputs over time, operating according to a certain reference time or activated by external events. Semantically, a constraint net represents a set of equations, with locations as variables and transductions as functions. The semantics of the constraint net, with each location denoting a trace, is the least solution of the set of equations. For trace and some other basic concepts of dynamic systems, the reader is referred to [10]. GivenCN , a constraint net model of a dynamic system, the abstract behavior of the system is the semantics of CN , denoted [[CN ]], i.e., the set of input/output traces satisfying the model. A complex system is generally composed of multiple components. A module is a constraint net with a set of locations as its interface. A constraint net can be composed hierarchically using modular and aggregation operators on modules. The semantics of a system can be obtained hierarchically from the semantics of its subsystems and their connections. A control system is modeled as a module that may be further decomposed into a hierarchy of interactive modules. The higher levels are typically composed of event-driven transductions and the lower levels are typically analog control components. The bottom level sends control signals to various e ectors, and at the same time, senses the state of sensors. Control signals ow down and state signals ow up. Sensing signals from the environment are distributed over levels. Each level is a grey box that represents the causal relationship between the inputs and the outputs. The inputs consist of the control signals from the higher level, the sensing signals from the environment and the current states from the lower level. The outputs consist of the control signals to the lower level and the current states to the higher level. 2.2 The CN Architecture of the Controller The soccer-playing softbot system is modeled as an integration of the soccer server and the controller (Fig. 1). The soccer server provides 22 soccer-playing 403 A Multi-level Constraint-based Controller for the Dynamo98 Robot Soccer Team